Classifying scanned documents is a challenging problem that involves image, layout, and text analysis for document understanding. Nevertheless, for certain benchmark datasets, notably RVL-CDIP, the state of the art is closing in to near-perfect performance when considering hundreds of thousands of training samples. With the advent of large language models (LLMs), which are excellent few-shot learners, the question arises to what extent the document classification problem can be addressed with only a few training samples, or even none at all. In this paper, we investigate this question in the context of zero-shot prompting and few-shot model fine-tuning, with the aim of reducing the need for human-annotated training samples as much as possible.
The advent of multimodal learning has brought a significant improvement in document AI. Documents are now treated as multimodal entities, incorporating both textual and visual information for downstream analysis. However, works in this space are often focused on the textual aspect, using the visual space as auxiliary information. While some works have explored pure vision based techniques for document image understanding, they require OCR identified text as input during inference, or do not align with text in their learning procedure. Therefore, we present a novel image-text alignment technique specially designed for leveraging the textual information in document images to improve performance on visual tasks. Our document encoder model DoPTA - trained with this technique demonstrates strong performance on a wide range of document image understanding tasks, without requiring OCR during inference. Combined with an auxiliary reconstruction objective, DoPTA consistently outperforms larger models, while using significantly lesser pre-training compute. DoPTA also sets new state-of-the art results on D4LA, and FUNSD, two challenging document visual analysis benchmarks.
Deep learning (DL) has revolutionized the field of document image analysis, showcasing superhuman performance across a diverse set of tasks. However, the inherent black-box nature of deep learning models still presents a significant challenge to their safe and robust deployment in industry. Regrettably, while a plethora of research has been dedicated in recent years to the development of DL-powered document analysis systems, research addressing their transparency aspects has been relatively scarce. In this paper, we aim to bridge this research gap by introducing DocXplain, a novel model-agnostic explainability method specifically designed for generating high interpretability feature attribution maps for the task of document image classification. In particular, our approach involves independently segmenting the foreground and background features of the documents into different document elements and then ablating these elements to assign feature importance. We extensively evaluate our proposed approach in the context of document image classification, utilizing 4 different evaluation metrics, 2 widely recognized document benchmark datasets, and 10 state-of-the-art document image classification models. By conducting a thorough quantitative and qualitative analysis against 9 existing state-of-the-art attribution methods, we demonstrate the superiority of our approach in terms of both faithfulness and interpretability. To the best of the authors' knowledge, this work presents the first model-agnostic attribution-based explainability method specifically tailored for document images. We anticipate that our work will significantly contribute to advancing research on transparency, fairness, and robustness of document image classification models.
This work explores knowledge distillation (KD) for visually-rich document (VRD) applications such as document layout analysis (DLA) and document image classification (DIC). While VRD research is dependent on increasingly sophisticated and cumbersome models, the field has neglected to study efficiency via model compression. Here, we design a KD experimentation methodology for more lean, performant models on document understanding (DU) tasks that are integral within larger task pipelines. We carefully selected KD strategies (response-based, feature-based) for distilling knowledge to and from backbones with different architectures (ResNet, ViT, DiT) and capacities (base, small, tiny). We study what affects the teacher-student knowledge gap and find that some methods (tuned vanilla KD, MSE, SimKD with an apt projector) can consistently outperform supervised student training. Furthermore, we design downstream task setups to evaluate covariate shift and the robustness of distilled DLA models on zero-shot layout-aware document visual question answering (DocVQA). DLA-KD experiments result in a large mAP knowledge gap, which unpredictably translates to downstream robustness, accentuating the need to further explore how to efficiently obtain more semantic document layout awareness.
This work addresses the need for a balanced approach between performance and efficiency in scalable production environments for visually-rich document understanding (VDU) tasks. Currently, there is a reliance on large document foundation models that offer advanced capabilities but come with a heavy computational burden. In this paper, we propose a multimodal early exit (EE) model design that incorporates various training strategies, exit layer types and placements. Our goal is to achieve a Pareto-optimal balance between predictive performance and efficiency for multimodal document image classification. Through a comprehensive set of experiments, we compare our approach with traditional exit policies and showcase an improved performance-efficiency trade-off. Our multimodal EE design preserves the model's predictive capabilities, enhancing both speed and latency. This is achieved through a reduction of over 20% in latency, while fully retaining the baseline accuracy. This research represents the first exploration of multimodal EE design within the VDU community, highlighting as well the effectiveness of calibration in improving confidence scores for exiting at different layers. Overall, our findings contribute to practical VDU applications by enhancing both performance and efficiency.
Zero-shot learning has been extensively investigated in the broader field of visual recognition, attracting significant interest recently. However, the current work on zero-shot learning in document image classification remains scarce. The existing studies either focus exclusively on zero-shot inference, or their evaluation does not align with the established criteria of zero-shot evaluation in the visual recognition domain. We provide a comprehensive document image classification analysis in Zero-Shot Learning (ZSL) and Generalized Zero-Shot Learning (GZSL) settings to address this gap. Our methodology and evaluation align with the established practices of this domain. Additionally, we propose zero-shot splits for the RVL-CDIP dataset. Furthermore, we introduce CICA (pronounced 'ki-ka'), a framework that enhances the zero-shot learning capabilities of CLIP. CICA consists of a novel 'content module' designed to leverage any generic document-related textual information. The discriminative features extracted by this module are aligned with CLIP's text and image features using a novel 'coupled-contrastive' loss. Our module improves CLIP's ZSL top-1 accuracy by 6.7% and GZSL harmonic mean by 24% on the RVL-CDIP dataset. Our module is lightweight and adds only 3.3% more parameters to CLIP. Our work sets the direction for future research in zero-shot document classification.
This paper proposes LayoutLLM, a more flexible document analysis method for understanding imaged documents. Visually Rich Document Understanding tasks, such as document image classification and information extraction, have gained significant attention due to their importance. Existing methods have been developed to enhance document comprehension by incorporating pre-training awareness of images, text, and layout structure. However, these methods require fine-tuning for each task and dataset, and the models are expensive to train and operate. To overcome this limitation, we propose a new LayoutLLM that integrates these with large-scale language models (LLMs). By leveraging the strengths of existing research in document image understanding and LLMs' superior language understanding capabilities, the proposed model, fine-tuned with multimodal instruction datasets, performs an understanding of document images in a single model. Our experiments demonstrate improvement over the baseline model in various document analysis tasks.
Digital libraries often face the challenge of processing a large volume of diverse document types. The manual collection and tagging of metadata can be a time-consuming and error-prone task. To address this, we aim to develop an automatic metadata extractor for digital libraries. In this work, we introduce the Heterogeneous Learning Resources (HLR) dataset designed for document image classification. The approach involves decomposing individual learning resources into constituent document images (sheets). These images are then processed through an OCR tool to extract textual representation. State-of-the-art classifiers are employed to classify both the document image and its textual content. Subsequently, the labels of the constituent document images are utilized to predict the label of the overall document.
This paper introduces a new large-scale dataset for Farsi document images, named SUT, which aims to tackle the challenges associated with obtaining diverse and substantial ground-truth data for supervised models in document image analysis (DIA) tasks, such as document image classification, text detection and recognition, and information retrieval. The dataset comprises 62,453 images that have been categorized into 21 distinct classes, including identity documents featuring synthetically generated personal information superimposed on various backgrounds. The dataset also includes corresponding files with labeling information for the images. The ground-truth data is organized in CSV files containing compiled image file paths and associated information about the embedded data. To demonstrate the efficacy of the SUT dataset in DIA tasks, it was utilized for document classification (achieving an accuracy of 86% using a convolutional neural network) and OCR (achieving a CER of 0.083 and 0.072 using Tesseract and EasyOCR engines, respectively). The SUT dataset represents a valuable resource for researchers who are interested in developing and evaluating supervised models in Farsi document image analysis.
Document image classification is different from plain-text document classification and consists of classifying a document by understanding the content and structure of documents such as forms, emails, and other such documents. We show that the only existing dataset for this task (Lewis et al., 2006) has several limitations and we introduce two newly curated multilingual datasets WIKI-DOC and MULTIEURLEX-DOC that overcome these limitations. We further undertake a comprehensive study of popular visually-rich document understanding or Document AI models in previously untested setting in document image classification such as 1) multi-label classification, and 2) zero-shot cross-lingual transfer setup. Experimental results show limitations of multilingual Document AI models on cross-lingual transfer across typologically distant languages. Our datasets and findings open the door for future research into improving Document AI models.
Visual document understanding (VDU) has rapidly advanced with the development of powerful multi-modal language models. However, these models typically require extensive document pre-training data to learn intermediate representations and often suffer a significant performance drop in real-world online industrial settings. A primary issue is their heavy reliance on OCR engines to extract local positional information within document pages, which limits the models' ability to capture global information and hinders their generalizability, flexibility, and robustness. In this paper, we introduce GlobalDoc, a cross-modal transformer-based architecture pre-trained in a self-supervised manner using three novel pretext objective tasks. GlobalDoc improves the learning of richer semantic concepts by unifying language and visual representations, resulting in more transferable models. For proper evaluation, we also propose two novel document-level downstream VDU tasks, Few-Shot Document Image Classification (DIC) and Content-based Document Image Retrieval (DIR), designed to simulate industrial scenarios more closely. Extensive experimentation has been conducted to demonstrate GlobalDoc's effectiveness in practical settings.
Visually-rich Document Understanding (VrDU) has attracted much research attention over the past years. Pre-trained models on a large number of document images with transformer-based backbones have led to significant performance gains in this field. The major challenge is how to fusion the different modalities (text, layout, and image) of the documents in a unified model with different pre-training tasks. This paper focuses on improving text-layout interactions and proposes a novel multi-modal pre-training model, LayoutMask. LayoutMask uses local 1D position, instead of global 1D position, as layout input and has two pre-training objectives: (1) Masked Language Modeling: predicting masked tokens with two novel masking strategies; (2) Masked Position Modeling: predicting masked 2D positions to improve layout representation learning. LayoutMask can enhance the interactions between text and layout modalities in a unified model and produce adaptive and robust multi-modal representations for downstream tasks. Experimental results show that our proposed method can achieve state-of-the-art results on a wide variety of VrDU problems, including form understanding, receipt understanding, and document image classification.
In the recent past, complex deep neural networks have received huge interest in various document understanding tasks such as document image classification and document retrieval. As many document types have a distinct visual style, learning only visual features with deep CNNs to classify document images have encountered the problem of low inter-class discrimination, and high intra-class structural variations between its categories. In parallel, text-level understanding jointly learned with the corresponding visual properties within a given document image has considerably improved the classification performance in terms of accuracy. In this paper, we design a self-attention-based fusion module that serves as a block in our ensemble trainable network. It allows to simultaneously learn the discriminant features of image and text modalities throughout the training stage. Besides, we encourage mutual learning by transferring the positive knowledge between image and text modalities during the training stage. This constraint is realized by adding a truncated-Kullback-Leibler divergence loss Tr-KLD-Reg as a new regularization term, to the conventional supervised setting. To the best of our knowledge, this is the first time to leverage a mutual learning approach along with a self-attention-based fusion module to perform document image classification. The experimental results illustrate the effectiveness of our approach in terms of accuracy for the single-modal and multi-modal modalities. Thus, the proposed ensemble self-attention-based mutual learning model outperforms the state-of-the-art classification results based on the benchmark RVL-CDIP and Tobacco-3482 datasets.
Adversarial attacks and defenses have gained increasing interest on computer vision systems in recent years, but as of today, most investigations are limited to images. However, many artificial intelligence models actually handle documentary data, which is very different from real world images. Hence, in this work, we try to apply the adversarial attack philosophy on documentary and natural data and to protect models against such attacks. We focus our work on untargeted gradient-based, transfer-based and score-based attacks and evaluate the impact of adversarial training, JPEG input compression and grey-scale input transformation on the robustness of ResNet50 and EfficientNetB0 model architectures. To the best of our knowledge, no such work has been conducted by the community in order to study the impact of these attacks on the document image classification task.
For many business applications that require the processing, indexing, and retrieval of professional documents such as legal briefs (in PDF format etc.), it is often essential to classify the pages of any given document into their corresponding types beforehand. Most existing studies in the field of document image classification either focus on single-page documents or treat multiple pages in a document independently. Although in recent years a few techniques have been proposed to exploit the context information from neighboring pages to enhance document page classification, they typically cannot be utilized with large pre-trained language models due to the constraint on input length. In this paper, we present a simple but effective approach that overcomes the above limitation. Specifically, we enhance the input with extra tokens carrying sequential information about previous pages - introducing recurrence - which enables the usage of pre-trained Transformer models like BERT for context-aware page classification. Our experiments conducted on two legal datasets in English and Portuguese respectively show that the proposed approach can significantly improve the performance of document page classification compared to the non-recurrent setup as well as the other context-aware baselines.
In this paper, we present StrucTexTv2, an effective document image pre-training framework, by performing masked visual-textual prediction. It consists of two self-supervised pre-training tasks: masked image modeling and masked language modeling, based on text region-level image masking. The proposed method randomly masks some image regions according to the bounding box coordinates of text words. The objectives of our pre-training tasks are reconstructing the pixels of masked image regions and the corresponding masked tokens simultaneously. Hence the pre-trained encoder can capture more textual semantics in comparison to the masked image modeling that usually predicts the masked image patches. Compared to the masked multi-modal modeling methods for document image understanding that rely on both the image and text modalities, StrucTexTv2 models image-only input and potentially deals with more application scenarios free from OCR pre-processing. Extensive experiments on mainstream benchmarks of document image understanding demonstrate the effectiveness of StrucTexTv2. It achieves competitive or even new state-of-the-art performance in various downstream tasks such as image classification, layout analysis, table structure recognition, document OCR, and information extraction under the end-to-end scenario.
In this paper, we propose to exploit the side-tuning framework for multimodal document classification. Side-tuning is a methodology for network adaptation recently introduced to solve some of the problems related to previous approaches. Thanks to this technique it is actually possible to overcome model rigidity and catastrophic forgetting of transfer learning by fine-tuning. The proposed solution uses off-the-shelf deep learning architectures leveraging the side-tuning framework to combine a base model with a tandem of two side networks. We show that side-tuning can be successfully employed also when different data sources are considered, e.g. text and images in document classification. The experimental results show that this approach pushes further the limit for document classification accuracy with respect to the state of the art.
Recent years have witnessed the rise and success of pre-training techniques in visually-rich document understanding. However, most existing methods lack the systematic mining and utilization of layout-centered knowledge, leading to sub-optimal performances. In this paper, we propose ERNIE-Layout, a novel document pre-training solution with layout knowledge enhancement in the whole workflow, to learn better representations that combine the features from text, layout, and image. Specifically, we first rearrange input sequences in the serialization stage, and then present a correlative pre-training task, reading order prediction, to learn the proper reading order of documents. To improve the layout awareness of the model, we integrate a spatial-aware disentangled attention into the multi-modal transformer and a replaced regions prediction task into the pre-training phase. Experimental results show that ERNIE-Layout achieves superior performance on various downstream tasks, setting new state-of-the-art on key information extraction, document image classification, and document question answering datasets. The code and models are publicly available at http://github.com/PaddlePaddle/PaddleNLP/tree/develop/model_zoo/ernie-layout.
Convolutional Neural Networks (ConvNets) have been thoroughly researched for document image classification and are known for their exceptional performance in unimodal image-based document classification. Recently, however, there has been a sudden shift in the field towards multimodal approaches that simultaneously learn from the visual and textual features of the documents. While this has led to significant advances in the field, it has also led to a waning interest in improving pure ConvNets-based approaches. This is not desirable, as many of the multimodal approaches still use ConvNets as their visual backbone, and thus improving ConvNets is essential to improving these approaches. In this paper, we present DocXClassifier, a ConvNet-based approach that, using state-of-the-art model design patterns together with modern data augmentation and training strategies, not only achieves significant performance improvements in image-based document classification, but also outperforms some of the recently proposed multimodal approaches. Moreover, DocXClassifier is capable of generating transformer-like attention maps, which makes it inherently interpretable, a property not found in previous image-based classification models. Our approach achieves a new peak performance in image-based classification on two popular document datasets, namely RVL-CDIP and Tobacco3482, with a top-1 classification accuracy of 94.17% and 95.57% on the two datasets, respectively. Moreover, it sets a new record for the highest image-based classification accuracy of 90.14% on Tobacco3482 without transfer learning from RVL-CDIP. Finally, our proposed model may serve as a powerful visual backbone for future multimodal approaches, by providing much richer visual features than existing counterparts.
Multimodal learning from document data has achieved great success lately as it allows to pre-train semantically meaningful features as a prior into a learnable downstream task. In this paper, we approach the document classification problem by learning cross-modal representations through language and vision cues, considering intra- and inter-modality relationships. Instead of merging features from different modalities into a joint representation space, the proposed method exploits high-level interactions and learns relevant semantic information from effective attention flows within and across modalities. The proposed learning objective is devised between intra- and inter-modality alignment tasks, where the similarity distribution per task is computed by contracting positive sample pairs while simultaneously contrasting negative ones in the joint representation space}. Extensive experiments on public document classification datasets demonstrate the effectiveness and the generality of our model on low-scale and large-scale datasets.